voxconverse icon indicating copy to clipboard operation
voxconverse copied to clipboard

labels for the test set

Open zhiyunfan opened this issue 4 years ago • 11 comments

The labels for the test set will be released to public after the VoxCeleb Speaker Recognition Challenge in October 2020. How can we download the labels for the test set? Looking forward to your reply.

zhiyunfan avatar Jul 01 '21 06:07 zhiyunfan

I'm interested in the labels for the test set as well

scalfs avatar Jul 12 '21 02:07 scalfs

The test set labels are now released.

JaesungHuh avatar Jul 25 '21 16:07 JaesungHuh

Thanks @JaesungHuh for sharing the test set labels.

For comparison with official results reported here, can you please confirm that these were computed on the subset of 232 files for which labels are available and not on the whole set of 312 files shared initially? This is important for the speaker diarization community to make sure we are not comparing apples and oranges.

cc @fnlandini @desh2608

hbredin avatar Jul 26 '21 14:07 hbredin

@hbredin Thanks for the question. Yes, the released Voxconverse test set are subset of 232 files from the whole set of 312 files shared initially. We did another few rounds of check to make labels more accurate and removed some files which annotators couldn't be 100% sure of their annotation. Please use this version from now on.

JaesungHuh avatar Jul 26 '21 14:07 JaesungHuh

Thanks for clarifying.

What should we call this version in publications: VoxConverse 2021 ? VoxConverse v0.0.2?

hbredin avatar Jul 26 '21 14:07 hbredin

I'll re-open this issue for other people to see. I have to discuss co-authors about this, but I think either is fine. Will let you know if the term fixed.

JaesungHuh avatar Jul 26 '21 15:07 JaesungHuh

We've recently released ver 0.3, fixing some of the errors in the test set labels. Please call "VoxConverse 0.3" when you use this dataset.

JaesungHuh avatar Jul 21 '22 13:07 JaesungHuh

Thanks for the heads-up @JaesungHuh.

Switching reference labels from 0.2 to 0.3 did "improve" my baseline by a whooping 2.8% (relative) in terms of speaker confusion rate. That is not negligible.

hbredin avatar Jul 22 '22 14:07 hbredin

Yes. We found these errors during the preparation for this year's VoxSRC workshop. I'll re-open this issue to let everyone know about this. I apologize for any inconvenience.

JaesungHuh avatar Jul 22 '22 15:07 JaesungHuh

Where can I find the link to the video files?

ahmadikalkhorani avatar Oct 26 '23 13:10 ahmadikalkhorani